After months of work, the two companies are making it straightforward for developers to run Meta's latest open-weight model on Apple hardware. Meta bills Muse Spark Open as its most capable open-weight release to date under a license that allows limited commercial use. It uses a mixture-of-experts approach and totals 816 billion parameters. Because it is open weight, developers can download the weights and run the model on a range of machines.
Apple and Meta say the model runs locally rather than in the cloud, which they argue adds a security layer and can lower costs. Customers will be able to connect between two and four Mac Studios to run Muse Spark for tasks like code generation, video analysis and other agentic work. Developers can download a version from sites such as Hugging Face that requires no extra configuration for Apple silicon.
"Muse Spark Open has also been tuned to run faster on Apple's hardware," and Apple shared fresh chip benchmarks along with technical documentation showing how its MLX framework can spread open-source models across multiple Macs. Meta AI chief Alexandr Wang put it bluntly: "We worked with Apple to bring leading Agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios, so you know we're really excited about this."
There is a catch: you need the fastest Apple chips and a lot of memory. A Mac Studio with the required M5 Ultra and 256GB of memory is at least $10,799, with a 512GB RAM configuration coming later this month. On top of that, users will need more than one machine. Earlier this year, Meta acquired hundreds of Mac computers, equipping the Superintelligence group to run experiments with on-device AI and agents.
This is a notable reset for two companies that spent years keeping each other at arm's length. Tensions spiked after Apple's 2021 privacy changes curtailed ad targeting; in early 2022, Meta estimated the hit to that year's revenue at $10 billion. The firms had worked together on iPhone integrations more than a decade ago, but they hadn't announced a public partnership since 2017, when Apple removed system-level Facebook support from iPhone.
The back-and-forth at the top has been sharp, too. During the Cambridge Analytica scandal in 2018, Tim Cook said, "I wouldn't be in this situation," and he frequently criticized Meta on privacy. Mark Zuckerberg told Joe Rogan last year that Apple "hasn't invented anything great since the iPhone," and he has long sought a platform Meta could control rather than rely on Apple.
Now, with hardware chief John Ternus succeeding Cook as Apple CEO last month, the tone is shifting.
When technology and markets intersect, it helps to stay informed, so sign up for the free Market Briefs to get daily clarity
After Zuckerberg posted on X, "Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple," Ternus replied, "Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!" Representatives for both companies declined to say whether this is the first step in a broader tie-up.
Wang said most of the animosity predated his time at Meta, and he pointed to former Apple director Daniel Gross, now on his team, as one reason the collaboration worked. Last year, Daniel Gross and ex-GitHub CEO Nat Friedman joined in headline-making hires as Zuckerberg reshaped Meta's AI strategy.
Wang said many powerful open-weight models are too large for most personal computers, but Mac Studio exposes enough memory to make Muse Spark Open practical. He also framed Muse Spark Open as a starter path for developers who might try it on a small setup and then scale to larger systems: "These aren't the most capable models that we've developed or released," he noted.
The move signals renewed faith in Meta's overhauled AI program and hints at Apple's growing clout in AI chips, even as Nvidia remains the data center heavyweight. Meta is balancing open-weight and proprietary work: recent milestones span the Muse personal agent application and, most recently, the proprietary foundation model Muse Wonder, which arrived earlier this week. Open source still figures in Zuckerberg's plans, too.
In August, Meta introduced a smaller open-weight model called Muse Glimmer, and he vowed that a larger and more capable release would follow. Meta had previously centered its AI strategy on the open-weight Llama family before shifting focus to proprietary models under Wang.
Apple, meanwhile, has been stacking model partnerships. In 2024 it revealed a deal with OpenAI to bring ChatGPT into Siri, which OpenAI later said in a filing tied to its legal dispute with Elon Musk's xAI "dramatically underperformed" and didn't boost user growth. In July, Apple sued OpenAI, accusing it of stealing trade secrets - an allegation OpenAI rejected. Earlier this year, Apple said it would use Google's Gemini models in Apple Intelligence and tap Google Cloud infrastructure.
Bottom line for regular users and builders: on-device AI is getting real, but it is not cheap. Running this locally points straight to premium hardware with serious memory and, in many cases, multiple Macs working together. That translates to bigger up front gear costs before you ever ship an AI feature.
If you want concise context around tech, finance and policy, subscribe to Market Briefs newsletter for five minute daily summaries
